Definition
A label is a target or reference value attached to an example for supervised learning, evaluation, or later analysis. It can be a class, number, span of text, ranking, bounding box, segmentation mask, preference, or structured answer.
A person may assign the label, or it may come from a measurement, policy rule, operational event, another model, or a combination of sources. "Known target" should not be read as unquestionable truth. Medical diagnoses change, fraud is discovered late, policies differ among reviewers, and model-generated labels can reproduce the model's own errors.
Label quality
The labeling protocol defines the task as much as the data does. Reviewers need instructions, examples, an allowed "uncertain" state, rules for disagreement, and evidence about inter-annotator consistency where judgment is involved. Majority vote can hide a genuinely ambiguous case.
Weak supervision uses noisy or indirect rules and signals to create labels at scale. Pseudo-labeling uses model predictions as temporary targets. Both can be useful when their uncertainty and provenance remain visible.
Distinguish it from nearby terms
- A feature is an input supplied to the model. A label is the target or reference answer.
- A prediction is the model's output. It may later become a pseudo-label, but the roles should not be confused during evaluation.
- An annotation is a recorded human or machine judgment. It becomes a label when the learning or evaluation task uses it as the target.
- Ground truth is a stronger claim that the reference accurately represents reality. Many labels are useful without meeting that standard.
Operational significance
Version labels with their source, time, policy, and adjudication history. Monitor class balance, disagreement, delayed outcomes, and changes in the process that produces the target. A model trained on consistent old policy may fail a current policy even when its code has not changed.
Check your understanding
Two reviewers disagree about whether a support ticket is abusive. The dataset should preserve or adjudicate that disagreement under a stated rule instead of treating one reviewer's answer as self-evident truth.